Anthropic’s new report highlights how cybercriminals are weaponizing its models

Anthropic has released a new report detailing "emerging trends" in how adversarial actors are misusing its Claude AI models, sounding alarms for the broader AI community.

While the company’s safeguards continue to block many harmful outputs, Anthropic acknowledged that threat actors are evolving their methods, using AI for increasingly complex and dangerous operations. The company shared the findings in a series of case studies, with the first one being a professional "influence-as-a-service" operation.

According to the report, actors leveraged Claude not just to create content but also to orchestrate bot activity across social media platforms like Twitter/X and Facebook. Claude helped these bots decide when to comment, like, or share posts, all tailored to political personas and designed for maximum influence. The operation engaged with tens of thousands of real users across multiple countries, although no single post achieved viral status.

"This is a distinct evolution in how influence operations are conducted," Anthropic said, noting that Claude served as an autonomous strategist for information campaigns, a significant shift from simple content generation.

Other cases detailed in the report include:

• Credential stuffing operations: Threat actors attempted to use Claude to enhance tools that scrape exposed usernames and passwords, particularly targeting internet-connected security cameras.

• Recruitment fraud campaigns: Scammers used Claude to polish scam communications aimed at job seekers in Eastern Europe, improving their credibility through real-time language refinement.

• Amateur malware creation: AIndividuals with limited technical skills used Claude to rapidly advance from basic scripting to developing sophisticated malware with features like facial recognition and dark web scanning.

Anthropic emphasized that the accounts involved were detected and banned in each instance. They also reported that there was no confirmed evidence that these efforts had been successfully deployed in the real world. Yet.

The report underscores a few troubling trends:

• Bad actors have started using AI to orchestrate complex abuse networks semi-autonomously.

• Generative models can drastically lower the skill barrier for cybercriminal activities.

Without solid safety systems, the misuse of frontier AI could significantly amplify threats across the digital ecosystem.

Anthropic credits its intelligence and monitoring programs, including tools like Clio and hierarchical summarization, with helping detect these activities early. They stressed the importance of continuous improvement in AI safety and called for closer collaboration between AI developers, governments, and the research community.

“We hope this report contributes to a broader understanding of the evolving threat landscape and strengthens collective defenses,” the company said.